Re: [Qemu-devel] Re: [PATCH 00/10] pci: pci_register_bar_simple

[Blue Swirl]

On Tue, Apr 5, 2011 at 12:26 AM, Michael S. Tsirkin <address@hidden> wrote:
> On Tue, Apr 05, 2011 at 12:05:08AM +0300, Blue Swirl wrote:
>> On Mon, Apr 4, 2011 at 10:44 PM, Michael S. Tsirkin <address@hidden> wrote:
>> > On Mon, Apr 04, 2011 at 08:02:23PM +0300, Blue Swirl wrote:
>> >> On Mon, Apr 4, 2011 at 7:35 PM, Avi Kivity <address@hidden> wrote:
>> >> > On 04/04/2011 07:22 PM, Anthony Liguori wrote:
>> >> >>
>> >> >> On 04/04/2011 10:59 AM, Michael S. Tsirkin wrote:
>> >> >>>
>> >> >>> On Mon, Apr 04, 2011 at 06:27:57PM +0300, Avi Kivity wrote:
>> >> >>>>
>> >> >>>> Many PCI BARs that use the memory address space map a single MMIO 
>> >> >>>> region
>> >> >>>> into
>> >> >>>> the entire BAR range.  Introduce an API pci_register_bar_simple() for
>> >> >>>> that use
>> >> >>>> case, and convert all users where this can be done trivially.
>> >> >>>>
>> >> >>>> This will reduce the work required to introduce a PCI memory API; 
>> >> >>>> it's
>> >> >>>> also
>> >> >>>> a nice code reduction in its own right.
>> >> >>>
>> >> >>> This will save some code, so
>> >> >>> Acked-by: Michael S. Tsirkin<address@hidden>
>> >> >>>
>> >> >>> I really hope the rest of devices will follow.
>> >> >>
>> >> >> How complete is this?
>> >> >
>> >> > I converted all devices which were easy to convert.  There may be one 
>> >> > or two
>> >> > more that can be converted with additional work (and perhaps with an
>> >> > additional pic_bar_get_current_address() API, and a 
>> >> > pci_bar_set_coalescing()
>> >> > API).  The rest likely need to stick with the callback-based API.
>> >>
>> >> In my version which I sent earlier but didn't commit, also other BARs
>> >> besides the first one and also tricky devices like VGA were handled.
>> >
>> > Yes, I liked that patchset too. What happened to it?
>>
>> Nothing, but I thought that there could be a "perfect" solution.
>>
>> I like in Avi's version that unnecessary API changes are avoided.
>
> Yes, it's nice that it's incremental.
>
>> >> But I didn't commit it because I felt it was not going to right
>> >> direction. I think the BARs should be specified in PCIDeviceInfo
>> >> instead of adding more function calls. The same applies to this patch
>> >> set.
>> >
>> > Is that really that fundamental? What I do care about is
>> > making pci.c track and register all device memory
>> > so that we can finally implement pci bridge features
>> > such as master abort handling and unmapped memory.
>>
>> The structure version can be done later. Right, pci.c should manage
>> the device mappings.
>
> OK, so applying Avi's patchset and building on that is
> your preferred approach too?

Avi's version is a bit too simple, at least multiple regions in a BAR
should be handled (for example macio.c needs that). But also that can
be added later (pci_register_bar_not_so_simple()?), so Avi's version
could be used as the starting point.